Keep characters recognizable across AI video scenes.
Completed owner study: 151 requests; 67 complete delivered clips visually reviewed and 84 failed requests. Reviewed 2026-10-03 (Europe/Sofia). Technical delivery does not establish character or scene fidelity.
Start with one approved character sheet. Reuse it unchanged, then write each scene separately. This can repeat a recognizable design, but cannot guarantee an identical face in every generation.
A prompt you can reuse
In the Character Generator, upload a picture you have permission to use or start with words. Review the visible traits, save the sheet, then change only the scene. Use “Review in studio” to check the prompt before generating.
FIXED CHARACTER Mara is a fictional adult, age 35, with a slender build, warm olive skin, a chin-length black bob with a left side part and one silver streak at her left temple. She has green almond-shaped eyes, a long oval face, high cheekbones, a narrow straight nose and a softly tapered jaw. She wears a plain navy wool coat, rust-red scarf, charcoal trousers and brown lace-up boots. SCENE At a quiet hotel reception, Mara opens a cream envelope and finds a brass key. She looks toward the closed door and says: “Who left this here?” CAMERA AND SOUND Natural daylight, realistic skin, a clearly visible face and quiet room ambience. SPOKEN LANGUAGE All dialogue is in English. Speak the written line exactly; do not translate it or add dialogue.
Save the character block. Change the scene and camera without changing the identity description. When starting from a picture, check the extracted eye color, hair length and small marks before saving. Remove temporary pose, lighting and background details. For French dialogue, explicitly request spoken French instead. Give each new request complete setting instructions: “same room” alone did not preserve our fight location.
Historical research, not supported portrait uploads
Current service limitation: face/person references are not supported for video generation. The portrait clips below are earlier research results, including refusals; they do not establish an available portrait feature. The character tool now transfers text descriptions only.
What our reviewed footage shows
| Method | Observed result |
|---|---|
| Fixed text | Hair and outfits repeat more reliably than face details. One mature-character pair has convincing broad resemblance; other pairs drift. |
| Photo → editable text | Several three-episode stories delivered from image-inspired text. Reviewing the extracted traits preserves more intended details, while fine faces and some clothes still drift. |
| Extra prose and centimeter targets | Paired clips still change the face. More detail has not established an exact identity lock. |
| Room and clothing images | Elena delivered three episodes with the intended broad design. Other clips retained room/outfit references but used the wrong character or repeated the wrong story beat. |
| Intact fictional portraits | Two original portraits delivered convincing source resemblance. Elena also delivered a second independent scene with convincing continuity, while an intervening same-image scene was refused for unverified-face protection. Two more Elena scenes subsequently delivered with convincing broad resemblance; one additional request was refused for its face reference and another failed because of rendering-account frequency limits. Acceptance remains inconsistent. |
| Several beats in one clip | Three connected beats can retain a coherent visible actor within one clip. Elena completed three settings; her fine face changed between them. |
| Two-character drama | A three-episode story delivered with distinct roles and recognizable styling. Fine faces, room design and some dialogue details drift. |
| Direct action wording | All three prompts in one fictional adult fight/stunt set delivered using explicit punch, kick and throw wording. Exact choreography varies; a fast throw distorted hair length. |
| Explicit spoken language | Three English controls delivered with the intended lines detected by automatic transcription/language checks. This is not verified lip-sync or a universal language guarantee. |
The Character Generator now offers optional connected-drama, dialogue and action starters. Choose one, edit the scene and language, and keep the saved character block unchanged. These starters adapt lessons from delivered footage; the edited examples themselves are not a guarantee of delivery or exact identity.
Blank-face and separate eye/nose/mouth references produced no usable video in four parts-based requests: one explicit face-protection refusal and three generation-voided failures without a specific cause. A complete-portrait control and a clothing-only control each delivered one clip. The parts method has no measured consistency result.
An intact-reference pair that delivered


Historical portrait experiment
These earlier experiments reused the same complete fictional portrait. This is not a supported customer workflow. For current text-only scenes, keep the appearance, hair, clothing and camera block unchanged; change only the scene. Our strongest set now includes four independently delivered clips, including the original pair. Face and wardrobe resemblance is convincing, while fine facial contours, room design and props still vary.
FIXED CHARACTER Mature adult, medium sturdy build. Fair skin with age lines and natural texture. Light brown hair with grey, tied in a high bun; short loose curls at the temples. Light blue eyes. Broad face, rounded jaw corners, narrow slightly convex nose, softly arched brows. Plain burgundy overcoat and dark slate-blue scarf. CAMERA Natural daylight, neutral white balance, realistic skin and anatomy. One actor. Begin with a well-lit frontal medium close-up, then a left profile. Keep age, hair, facial structure and wardrobe unchanged. SCENE Write the new scene here. Keep the setting and lighting instructions complete.
The latest four-request follow-up delivered two complete 30-second clips and failed twice: one face-reference refusal and one rendering-account frequency restriction. No failed input was resubmitted. This establishes useful resemblance, not reliable acceptance or perfect consistency.


Compare the face separately from the costume


Check the face at the start, middle and end, plus frames during fast motion. Compare earlier scenes under similar lighting. Check dialogue and action separately too: a finished video can miss a requested line or story beat.
Original fictional adult characters; actual complete decoded files; six regular sampled frames per clip, supplemented by action and dialogue-timing checks; unblinded visual review. Broad resemblance is not biometric identity verification. Automatic transcripts do not verify lip-sync. Three Elena detailed/maximum pairs used identical inputs and are repeated draws, not separate detail levels. These are owner experiments, not customer fulfillment statistics.
Research behind the tests
CellCog's firsthand workflow describes fixed text with scenery and clothing references. Our intact-portrait successes show its blanket face-rejection claim does not describe every result in our tests. One creator's small portrait test reports one acceptance and four refusals, plus better continuity inside one clip. That is a self-reported experiment on another route, not our acceptance rate.